OpenAI Restricts GPT-5.6 to Trusted Partners as US Government Tightens AI Model Oversight

On June 26, 2026, OpenAI officially previewed the GPT-5.6 series: flagship model Sol, balanced model Terra, and fast, cheap model Luna. The series surpasses Anthropic's Claude Mythos 5 on coding benchmarks, marking OpenAI's return to the performance throne.

Pricing is tiered by capability: Sol at $5 input/$30 output per million tokens, Terra at $2.5 input/$15 output, and Luna at $1 input/$6 output. Compared to GPT-5.5, Terra delivers equivalent performance at half the cost. Additionally, GPT-5.6 introduces more predictable prompt caching: support for explicit cache breakpoints and a 30-minute minimum cache lifetime. Cache writes are billed at 1.25x the model's non-cached input rate, and cache reads receive a 90% discount.

OpenAI disclosed in the announcement that, as part of ongoing engagement with the U.S. government, they previewed the models and capabilities before release. At the government's request, OpenAI initially granted access only to a small set of trusted partners, and the list of these partners has been shared with the government. This practice goes beyond standard pre-release briefings, effectively giving the government veto power over the deployment of next-generation AI models.

The Washington Post reported that the U.S. government will decide who can use GPT-5.6. This description is corroborated by OpenAI's statement mentioning "trusted partners" with "participation shared with the government." On Hacker News, 357 users engaged in heated discussion on the topic.

The Decoder directly quoted OpenAI's view in its report: the company considers government-led access restrictions "unsustainable." This phrasing appears in the context of the official announcement, implying that while OpenAI complied with the request, it is now publicly expressing discontent.

Notably, OpenAI used the phrase "limited preview" rather than "formal release" in the announcement and promised "broad availability within weeks." This suggests government intervention may be temporary, but it could also set a precedent for future models. Anthropic's Claude Mythos 5 faces a similar review environment—both companies released models on the same day (as noted by Latent Space: "OAI and ANT on the same day with oddly tiered releases"), hinting at unified control by the government over multiple frontier models.

GPT-5.6 Sol beats Claude Mythos 5 on coding benchmarks, indicating that whether from a capability or commercial competition standpoint, it represents the highest level of current AI technology. But it is precisely this capability that has raised government concerns.

One consequence of the limited preview is that the most advanced tools are initially held only by government-approved entities. While this may reduce the risk of misuse, it also exacerbates power concentration—governments and companies jointly decide who can leverage the strongest AI. OpenAI describes this model as "unsustainable," likely anticipating backlash from other developers, researchers, and competitors.

The release pattern of GPT-5.6 may become a watershed moment for AI governance. Although the U.S. government has not directly regulated AI, it has informally controlled model distribution through "requests." OpenAI's public complaint indicates that even the largest AI companies feel the pressure of such intervention.

If other governments follow suit, future frontier models may face a fragmented landscape of multi-country approvals. OpenAI repeatedly emphasizes its commitment to "broad accessibility" in the announcement, but the current reality is that a small circle of government and trusted partners will get early experience before full public release.

This event also highlights the tension between safety and openness: the government fears model capabilities could be misused, while OpenAI worries that excessive restrictions could harm innovation and commercial returns. Ultimately, who decides "who gets it first and who gets it later" will redefine the power structure of the AI industry.